Sensors 2009, 9, 7753-7770; doi:10.3390/s91007753 OPEN ACCESS
sensors ISSN 1424-8220 www.mdpi.com/journal/sensors Article
Defect Detection in Arc-Welding Processes by Means of the Line-to-Continuum Method and Feature Selection P. Beatriz Garcia-Allende *, Jesus Mirapeix, Olga M. Conde, Adolfo Cobo and Jose M. Lopez-Higuera Photonics Engineering Group, University of Cantabria, Avda. de los Castros S/N, 39005 Santander, Spain; E-Mails:
[email protected] (J.M.);
[email protected] (O.M.C.);
[email protected] (A.C.);
[email protected] (J.M.L.-H.) * Author to whom correspondence should be addressed; E-Mail:
[email protected]; Tel.: +34-942-200877; Fax: +34-942-200877. Received: 8 July 2009; in revised form: 28 August 2009 / Accepted: 24 September 2009 / Published: 29 September 2009
Abstract: Plasma optical spectroscopy is widely employed in on-line welding diagnostics. The determination of the plasma electron temperature, which is typically selected as the output monitoring parameter, implies the identification of the atomic emission lines. As a consequence, additional processing stages are required with a direct impact on the real time performance of the technique. The line-to-continuum method is a feasible alternative spectroscopic approach and it is particularly interesting in terms of its computational efficiency. However, the monitoring signal highly depends on the chosen emission line. In this paper, a feature selection methodology is proposed to solve the uncertainty regarding the selection of the optimum spectral band, which allows the employment of the line-tocontinuum method for on-line welding diagnostics. Field test results have been conducted to demonstrate the feasibility of the solution. Keywords: arc-welding; plasma spectroscopy; feature selection; on-line monitoring
Sensors 2009, 9
7754
1. Introduction The fields where welding processes are key factors in the production stages cover a wide range of different applications: manufacturing of pipes for various sectors, engines for aeronautics, automobiles or heavy components for nuclear power stations are some relevant examples in this regard. The lack of a complete comprehension of the physical phenomena occurring during the welding process, and the demanding quality standards to be found in this framework, have forced scientists to carry out an intense research effort in both welding physics and procedures devoted to cope with quality issues. Some of these studies have been focused on the development of theoretical models for both arc and laser welding [1-3], including numerical analysis approaches [4]. These efforts help to understand the process and, therefore, to determine the precise input parameter ranges that will provide seams free of flaws. However, in practice welding coupons employed for parameter adjustment, and both destructive and non-destructive trials [5] have to be used to ensure that the performed seams satisfy the established quality standards. This obviously implies a significant cost in terms of productivity, as a lot of time is spent before and after the welding process itself, and, therefore, some of the seams have to be reworked and evaluated again. This scenario has led to an intense research effort aimed at developing efficient and reliable on-line welding quality monitoring systems. They should be able to detect in real-time the occurrence of possible defects and, as an added value, to control the welding setup to try to avoid these defects or drifts from the standard operation conditions. Several techniques have been proposed, from electrical and capacitive sensors [6,7], to monitoring based on the analysis of the acoustic signal generated during the process [8,9] or solutions based on machine vision [10,11]. Among these alternatives, the optical analysis of the welding plasma radiation has proved to be a feasible and promising option. Initial proposals were based on the use of photodiodes and the analysis of emissions in the ultraviolet, visible and infrared regions [12], determining for example the full-penetration condition in laser welding [13]. A more sophisticated approach has been proposed by considering plasma optical spectroscopy, where emission lines appearing in the plasma spectra are analyzed to provide a plasma electron temperature Te profile that shows a direct correlation to weld defects [14,15]. In the last years, several publications have dealt with refinements of this technique, allowing automatic defect detection [16] and reducing the overall computational cost of the system [17]. More recently, new strategies have been proposed to extract more information from the plasma spectra, like the correlation analysis proposed by Sibillano et al. [18], or proposals based on the use of optimization algorithms to generate synthetic spectra [19]. Within the same framework, new spectroscopic parameters are also being studied in an attempt to improve the monitoring system efficiency [20]. One of the key issues when using plasma spectroscopy lies in the correct selection of the emission lines chosen to calculate the output monitoring parameter. On the one hand, and depending on the selected instrumentation, there can be ambiguities on the emission line identification, what can end in unexpected results. On the other hand, and especially when defect classification is required, i.e., to be able to distinguish among different types of defects, it would be highly interesting to know which emission lines allow a better discrimination for classification purposes.
Sensors 2009, 9
7755
We have conducted some previous studies by using PCA (Principal Component Analysis) and SFFS (Sequential Forward Floating Selection) to feed an Artificial Neural Network [21,22]. The use of SFFS allows to gain knowledge about the best spectral bands selected. This will be used in this paper to propose a scheme based on both the SFFS algorithm and the line-to-continuum method [23] to generate the required output monitoring profiles. The line-to-continuum method implies the use of only a single emission line that, in addition, does not need to be identified, i.e., associated with its chemical species. 2. Plasma Optical Spectroscopy for Welding Diagnostics The plasma electron temperature has been widely used as the output monitoring parameter for welding diagnostics, given the known correlation between its profiles and the appearance of defects in the seams. There are basically two approaches that are employed in the literature: a precise estimation of Te can be obtained with the Boltzmann-plot method [23]: ⎛I λ ln⎜⎜ mn mn ⎝ Amn g m
⎞ hcN ⎞ E m ⎟⎟ = ln⎛⎜ ⎟− ⎝ Z ⎠ kTe ⎠
(1)
where several emission lines from the same species are involved in the calculations. In the previous equation Imn is the relative intensity of the chosen emission line, m and n the upper and lower states, respectively, λmn the central wavelength associated with the line, Amn the transition probability, gm the statistical weight, h the Planck’s constant, c the light velocity, N the population density of the state m, Z the partition function, Em the upper level energy and k the Boltzmann constant. Te can be obtained if the left-hand side of Equation (1) is represented versus Em, given that the slope of the resulting line is inversely proportional to the temperature. On the other hand, and due to considerations regarding the computational performance of the monitoring system, which determines its spatial resolution, a simplification of the Boltzmann-plot method, where only two emission lines are involved, is typically used: Te =
Em (2) − Em (1) ⎡ E (1) I (1) A(2) g m (2)λ (1) ⎤ k ln ⎢ m ⎥ ⎣ Em (2) I (2) A(1) g m (1)λ (2) ⎦
(2)
This equation was proposed by Marotta [24] for arc-welding processes. The commented techniques can be equally applied to both arc and laser processes, although for the latter the energies of the upper level disappear from the logarithm in the denominator. Although different approaches can be taken into account, a possible processing scheme designed to provide the required Te estimation from the acquired welding plasma spectra is presented in Figure 1. The identification of the emission lines is compulsory to obtain Te. As shown in Figure 1, this requires three additional processing stages (peak detection and line modeling and identification) and it has, as a consequence, a direct implication in the real-time performance of the overall approach. An alternative solution is to perform a previous spectral band selection stage with a data set consisting of spectra from the same welding process under different conditions. Afterwards those lines can be used without involving the identification in the processing scheme.
Sensors 2009, 9
7756
Figure 1. Comparison of the processing schemes associated with the traditional approach and the proposed solution based on SFFS and the line-to-continuum method.
This could be applied for scenarios where the same materials and welding conditions are used, but it limits the flexibility of the analysis strategy. On the other hand, as previously commented, there is a lack of knowledge on the selection of the optimal emission lines for welding diagnostics. Some studies have been carried out comparing the response of different elements and species, but we believe that by specifically searching for the most discriminant spectral bands the overall performance of the monitoring system should be improved. Within this framework, the use of the line-to-continuum method to generate the output monitoring profiles seems to be a good solution, given that it does not require the identification of the chosen emission line. However, it could be performed to avoid problems related to the effect of unresolved lines. This method was originally intended to estimate Te by means of the following expression [25]:
εl Ic
(λ ) = 2.0052 × 10 −5
⎛ E − Em Amn g m 1 exp⎜⎜ i Z i Te ξ ⎝ kTe
⎞ λ ⎟⎟ ⎠ Δλ
(3)
where εl is the line intensity integrated over the line profile, Ic the intensity of the adjacent background radiation (non-integrated), Zi is the ion partition function, ζ the free-bound continuum correction, Ei the ionization potential and Δλ the wavelength bandwidth. It is worth mentioning that an iterative method has to be employed to determine Te via Equation (3). However, in the proposed method we only use the εl / Ic ratio as the monitoring parameter. In a previous paper, this approach was initially explored in comparison to an alternative method based on the estimation of the wavelength associated with the maximum intensity of the continuum radiation [20]. In this case we concluded that the line-tocontinuum method could not be reliably used given the uncertainty regarding the selection of the spectral band for the subsequent analysis. In this paper the use of the SFFS algorithm will help to deal with this issue.
Sensors 2009, 9
7757
3. Sequential Floating forward Selection of Spectral Bands The Sequential Floating Forward Selection (SFFS) algorithm [26] is widely applied to reduce the dimensionality, i.e., the number of features or wavelengths, of spectral data prior to their interpretation [27,28]. The spectral band selection criterion is based on the capability of the distinct features to separate the different classes to be discriminated afterwards. The greater the separability between the classes that a wavelength provides, the better the wavelength is for classification purposes. Therefore, a similar approach could be followed to solve the uncertainty encountered in the selection of the optimum band within the plasma spectra for on-line welding quality monitoring by means of the line-to-continuum method described above [20]. The aim of SFFS is to select M spectral bands that best discriminate among correct welds and flaws, out of the total number of initial bands N, so M
